哲学家兼 AI 研究员 Murray Shanahan 探讨 AI 引发的哲学问题、科幻预测的准确性以及 John McCarthy 的遗产。
Philosopher and AI researcher Murray Shanahan discusses the philosophical questions raised by AI, the accuracy of sci-fi predictions, and the legacy of John McCarthy.
要点 · TL;DR
智能与意识可分离；AI 可以智能而无意识。 Intelligence and consciousness are dissociable; AI can be intelligent without being conscious.
具身性是人类认知的关键；大语言模型缺乏它，限制了深层理解。 Embodiment is key to human cognition; LLMs lack it, limiting deep understanding.
我们需要为 AI 创造新语言，如‘异类心智实体’，以避免误导性的拟人化。 We need new language for AI, like 'exotic mindlike entities', to avoid misleading anthropomorphism.
核心观点 · Key points
智能与意识是可分离的；我们可以拥有没有意识的智能系统。 Intelligence and consciousness are dissociable; we can have intelligent systems without consciousness.
具身性是人类认知的基础；语言模型缺乏这一点，限制了深层理解。 Embodiment grounds human cognition; language models lack this, limiting deep understanding.
图灵测试是衡量 AI 能力的糟糕标准；它忽略了具身性和更深层的认知。 The Turing test is a poor measure of AI capability; it ignores embodiment and deeper cognition.
拟人化 AI 是自然的但有风险；可能误导对其真实能力的认知。 Anthropomorphizing AI is natural but risky; it can mislead about true capabilities.
我们需要新的语言来描述 AI，例如“异类心智类实体”，以避免高估或低估。 We need new language for AI, like 'exotic mindlike entities', to avoid over- or under-estimation.
反共识 · Contrarian takes
大型语言模型能在日常意义上推理，而不仅仅是模仿训练数据。 Large language models can reason in an everyday sense, not just mimic training data.
意识不是二元的；它有多个方面，可以在 AI 中分离。 Consciousness is not binary; it has multiple facets that can be dissociated in AI.
加兰测试（看已知机器人是否被认为有意识）比图灵测试更相关。 The Garland test (seeing if a known robot is considered conscious) is more relevant than Turing test.
对 AI 礼貌（说请/谢谢）会因角色扮演模仿而改善回应。 Being polite to AI (saying please/thank you) improves responses due to roleplaying mimicry.
未来 AI 可能因叙事和接触而被视为有意识，就像章鱼一样。 Future AI may be treated as conscious based on narratives and encounters, like octopuses.
本期章节 · Chapters（共 21）
AI 的哲学问题Philosophical questions about AI
《机械姬》与《她》的预测Ex Machina and Her predictions
约翰·麦卡锡与“人工智能”术语John McCarthy and the term 'artificial intelligence'
符号 AI 与现代方法对比Symbolic AI vs modern approaches
符号 AI 及其局限Symbolic AI and its limitations
LLM 回归符号思想Return to symbolic ideas with LLMs
AI 推理能力与符号系统比较AI's reasoning ability compared to symbolic systems
论 LLM 中的推理On reasoning in LLMs
图灵测试及其局限Turing test and its limitations
具身智能与空间隐喻Embodied intelligence and spatial metaphors
图灵测试替代：加兰测试Alternatives to Turing test: Garland test
加兰测试：意识与智能The Garland Test and Consciousness vs. Intelligence
拟人化与民间心理学Anthropomorphization and Folk Psychology
通过语言拟人化 AIAnthropomorphizing AI through language
具身 AI 与智能Embodied AI and intelligence
智能与意识的分离Dissociating intelligence from consciousness
AI 中的意识与自我意识Awareness and Self-Awareness in AI
AI 中的情感痛苦与涌现Emotional Suffering and Emergence in AI